AI agents for business, with people in charge
AI agents are software systems that use a large language model to plan and complete multi-step tasks, calling your tools and data within limits you set. Arke Technologies, in Johannesburg, designs, builds and governs AI agents for South African organisations, public bodies and schools, with a person approving every action that matters.
What are AI agents?
An AI agent is software given a goal rather than a script. It uses a large language model to work out the steps, calls tools such as your email, ticketing or finance system to carry them out, checks the result and decides what to do next. A chatbot answers; an agent acts.
This is a practical introduction to AI agents for the people who must approve them, rather than a definitive guide to every framework. AI agents use context to decide: the request in front of them, the records they may read and the rules you set. They do not learn on their own between tasks, and they can be wrong with confidence. So the design question is always what the agent may do alone, and what it must propose to a person.
Agents are part of our AI consulting and implementation service. For how agents differ from the models behind them, read AI agents vs generative AI.
AI automation and AI agents: which do you need?
Most organisations need AI automation and AI agents in different places. The difference between AI agents and workflow automation is who writes the steps: in a workflow your team fixes the sequence and AI handles single steps; an agent chooses its own sequence within limits. Automated AI agents suit tasks whose path varies case by case. Fixed workflows suit the rest, because they are cheaper to run and simpler to audit. If you can draw the process on one page, start with AI automation and workflow automation.
The AI agents tech stack we build on
Models. Claude from Anthropic, Gemini through Vertex AI, or an open-weight model, chosen per task on accuracy, cost and where data may be processed.- Tools. API connections to your systems from our API integration team, each with the least access that does the job.
- Knowledge. Retrieval over approved documents, so answers cite a source and hallucinations are easier to catch.
- Controls. Approval points, step and spending limits, and filtering against prompt injection.
- Records. Logs of every prompt, tool call, approval and output.
We build so the model can be swapped without rebuilding the agent.
Why AI agents fail, and how we prevent it
Separating value from hype is most of the work. A demo at a hackathon or a conference runs on clean data and a patient audience; production has neither. The failures we design against follow a pattern:
- Too much autonomy too soon. We start in draft mode and extend autonomy only to proven actions.
- Hallucinations. We ground agents in approved sources, require citations and route low-confidence outputs to a person.
- Poor data and weak integrations. We check data readiness first and use tested APIs, not screen-scraping.
- Prompt injection. A malicious email can instruct an agent to leak data, so access is least-privilege and our cybersecurity team tests it.
- Runaway cost. Agents loop, so we set step limits and budgets and report spend monthly.
- No owner. We agree a named process owner on each side.
Models and tools will keep improving, but the growth an organisation gets from agents in future depends on governance: an agent is given more autonomy only when its record on your own cases shows it has earned it.
Governance, compliance, procurement and continuity
- Governance and reporting. A named service lead; a monthly report on sampled accuracy, approvals, overrides, cost and incidents; escalation paths; change control for prompts, tools and model versions. Our AI governance framework sets the approval points: AI recommends, people decide, every action is logged.
- Compliance. POPIA controls, including section 71 limits on decisions based solely on automated processing; encryption, access controls and data minimisation; ISO 27001-aligned security; King IV Principle 12 reporting; PFMA or MFMA documentation for public bodies. We never train models on client data without explicit consent. See POPIA compliance.
- Procurement. Level 1 B-BBEE contributor, giving 135% procurement recognition (sworn affidavit available on request); company documents, bank confirmation letter, CSD supplier number and SARS tax compliance status on request; standard bidding documents (SBD forms) completed when a bid requires them. Buy by proposal, SLA and purchase order, or by tender; see procuring from Arke.
- Continuity. You own the prompts, tool definitions, test sets, logs, code and documentation, and on exit we hand over everything, credentials included.